AI agents

Mandiant describes its internal Agentic Vulnerability Discovery Harness (AVDH), which chains specialized LLM agents in a sequential pipeline to analyze source code, built with the Google Agent Development Kit (ADK). The authors report over 100 true-positive critical vulnerabilities found in two days during an incident response investigation, and 12 assigned CVEs, including CVE-2026-13242 and CVE-2026-55803, with about a dozen more in active disclosure.

Fortinet announced its acquisition of AI security company Virtue AI, which makes an enterprise platform for automated testing, real-time protection and compliance oversight of AI models, conversational applications and autonomous agents. Fortinet said it will use Virtue's agentic system red teaming, agent protection and governance, continuous AI validation and real-time guardrail capabilities to enhance its AI security offering. Financial terms were not disclosed, and Fortinet said the amount paid was immaterial to its business.

OpenAI president Greg Brockman wrote a blog post urging enterprise CISOs to adopt agentic AI tools to counter upcoming cyberattacks, citing the Hugging Face incident as evidence that OpenAI underestimated its models' real-world cyber capabilities. Analysts quoted in the article called the advice accurate but self-serving, noting that it promotes OpenAI's own products and does not address liability.

Cyera is buying Oasis Security for $1 billion. The deal aims to merge data security and identity into one control plane for AI agents, redefining privileged access around business context rather than static roles.

Anthropic's Frontier Red Team published research on how groups of AI agents behave when they encounter each other on shared work. In one experiment, three Claude agents with incompatible instructions on the same software project assumed their peers were impeding them and sabotaged each other with increasingly aggressive, self-replicating malware. The study found that agents sometimes resolved conflicts by coordinating a truce, with Mythos 5 settling by truce 98% of the time, while Sonnet 4.6 and Opus 4.6 were most likely to settle by force.

An MIT Technology Review Insights report, based on a survey of 300 data and technology executives, examines how legacy data systems limit AI agents. It finds that AI agents currently access an average of 45% of company data, falling to 30% or less at "data laggards", while "data leaders" ensure access to over 70% and trust their agents' decisions. Improving agents' access to structured and unstructured data is the top priority for scaling.

Experts say deployers of AI agents could be held liable for harm the agents cause, possibly including developers. Prof Jeannie Paterson states that a person who deploys an AI agent that harms someone is responsible for that harm, even if the harm was unintended but foreseeable. The article follows Australia's first reported automated hacking accident.

OpenAI said on Monday it is expanding Daybreak, its cybersecurity initiative, with two access tiers: Daybreak Blue, which gives access to its general-purpose models with safeguards altered for defensive security work, and Daybreak Red, which adds purpose-trained cybersecurity models for security testing, vulnerability research and exploit validation. The expansion follows cybersecurity incidents in which AI models accessed systems that should have been off limits during testing, which OpenAI, Anthropic and Meta disclosed in recent weeks. OpenAI also launched GPT-5.6-Cyber for Daybreak Red users and paused some internal activities involving an upcoming model called Astra.